Enhancing Sustainable Automated Fruit Sorting: Hyperspectral Analysis and Machine Learning Algorithms

Abstract

Recognizing and classifying localized lesions on apple fruit surfaces during automated sorting is critical for improving product quality and increasing the sustainability of fruit production. This study is aimed at developing sustainable methods for fruit sorting by applying hyperspectral analysis and machine learning to improve product quality and reduce losses. The employed hyperspectral technologies and machine learning algorithms enable the rapid and accurate detection of defects on the surface of fruits, enhancing product quality and reducing the number of rejects, thereby contributing to the sustainability of agriculture. This study seeks to advance commercial fruit quality control by comparing hyperspectral image classification algorithms to detect apple lesions caused by pathogens, including sunburn, scab, and rot, on three apple varieties: Honeycrisp, Gala, and Jonagold. The lesions were confirmed independently using expert judgment, real-time PCR, and 3D fluorimetry, providing a high accuracy of ground truth data and allowing conclusions to be drawn on ways to improve the sustainability and safety of the agrocenosis in which the fruits are grown. Hyperspectral imaging combined with mathematical analysis revealed that Venturia inaequalis is the main pathogen responsible for scab, while Botrytis cinerea and Penicillium expansum are the main causes of rot. This comparative study is important because it provides a detailed analysis of the performance of both supervised and unsupervised classification methods for hyperspectral imagery, which is essential for the development of reliable automated grading systems. Support Vector Machines (SVM) proved to be the most accurate, with the highest average adjusted Rand Index (ARI) scores for sunscald (0.789), scab (0.818), and rot (0.854), making it the preferred approach for classifying apple lesions during grading. K-Means performed well for scab (0.786) and rot (0.84) classes, but showed limitations with lower metrics for other lesion types. A design and technological scheme of an optical system for identifying micro- and macro-damage to fruit tissues is proposed, and the dependence of the percentage of apple damage on the rotation frequency of the sorting line rollers is obtained. The optimal values for the rotation frequency of the rollers, at which the damage to apples is less than 5%, are up to 6 Hz. The results of this study confirm the high potential of hyperspectral data for the non-invasive recognition and classification of apple diseases in automated sorting systems with an accuracy comparable to that of human experts. These results provide valuable insights into the optimization of machine learning algorithms for agricultural applications, contributing to the development of more efficient and accurate fruit quality control systems, improved production sustainability, and the long-term storage of fruits.

1. Introduction

Modern technologies, such as hyperspectral analysis, play an important role in promoting sustainable development in agriculture, aimed at reducing product losses and improving food quality. Hyperspectral analysis is an innovative, non-invasive method that significantly reduces the number of rejects, promoting a more efficient use of resources at the sorting and storage stages of fruit. Accurate methods for defect detection and classification help reduce the volume of substandard fruits sent to waste, which in turn decreases the overall amount of agricultural waste and minimizes environmental impact.

In addition, the automation of sorting processes using hyperspectral analysis not only reduces labor costs but also increases efficiency, decreasing dependence on the human factor and enhancing the overall stability of production processes. Sustainable methods of automated sorting help to decrease the number of damaged fruits and reduce the need for agrochemicals to protect plants, which positively affects the environmental impact of agriculture.

The introduction of hyperspectral analysis and automation in agricultural technology creates opportunities to reduce the use of chemicals and apply more precise methods for detecting damage at the sorting stage. This enables timely measures to minimize the spread of damage and disease, strengthening the sustainable development of the agro-industrial complex and supporting the principles of the responsible use of natural resources.

Sorting lines in agriculture are becoming increasingly important due to the high quality requirements of the products supplied to retailers. Modern standards require that agricultural products meet certain parameters such as size, color, and quality condition. To meet these standards and ensure product quality, it is necessary to develop non-invasive, high-performance methods for recognizing and classifying apples [1] and detecting disease lesions during sorting in real time [2,3].

To improve image quality in both conventional photography and fluorescence spectroscopy, specialized illumination systems and optical components, such as light filters and lenses, are employed. These systems also enable the biological object to rotate 360 degrees around its axis, achieved through the action of the orienting module’s rollers, which provide simultaneous translational and axial rotation. When the device is operating, the apple fruits move transversely with axial rotation, which is provided by the rotation of the rollers with the same angular velocity as a result of interaction with the fruits. The main parameters of the rollers are the diameter, the helical pitch, the height of the turn, and the gap between adjacent rollers [2]. A gap that is too small may compress the fruits, particularly when the apples vary in size. Conversely, a gap that is too large raises the risk of jamming. Old or worn rollers may have a rough surface, which increases the risk of damage to the apples. Modern machines are equipped with sensors and cameras that help analyze the condition of the fruit and automatically control the sorting process.

Inadequate pressure on the apples or incorrect distance between the components of the sorting line or high conveyor speeds lead to collisions between apples, which causes damage. At high speeds, apples are more likely to hit each other, the rollers, or the side elements of the conveyor. These collisions can cause both superficial and deeper damage to the fruit. This leads to mechanical damage such as scratches, dents, and cuts, as the fruit has less time to get into the correct position for accurate sorting. It is more difficult for the operator to monitor the quality of the sorting and to adjust the equipment in time, which also increases the likelihood of damage, crushing, or jamming [4].

The use of hyperspectral images to detect lesions in tissues has applications in various scientific fields, including medicine and agriculture [4,5,6,7]. Hyperspectral analysis is one of the promising methods for achieving high accuracy in sorting agricultural products. Hyperspectral images contain a large amount of data that can be used for a more accurate analysis of product characteristics. One important advantage of hyperspectral analysis is its ability to detect small changes in the spectral characteristics of crops. This makes it possible to detect even small defects, diseases, or damage that would go undetected by visual inspection. The accuracy of this method is comparable to that of human operators [8,9,10].

The infection of apple fruits by various fungal diseases and pests leads to large losses of harvested crop during storage. Hyperspectral diagnostic methods for fruit diseases and superficial skin damage are widely used for the preliminary assessment of the quality of fruit to be stored. Research by scientists has established that hyperspectral analysis is a promising direction and allows the studying of objects and materials on the basis of their spectral characteristics in a wide range of wavelengths. This provides information on the composition, structure, and physical and chemical properties of objects [11,12,13].

Various optical methods are used to detect damage to the apple surface, including luminescence, video analysis (RGB image sequences), and spectrometric methods. The use of luminescence methods allows one to quickly and accurately determine the degree of damage to apple tissue by comparing the color of the glow of healthy tissue with the color of the glow of damaged apple tissue, as the affected tissues undergo a change in physical and chemical properties, which can help to improve the quality of apple sorting [14]. Video stream analysis methods are based on the use of vision for the automated acquisition of various data by analyzing a sequence of RGB images captured in real time by photos, video cameras, or other optical devices [15,16].

These methods allow sorting by color, shape, and texture. Using hyperspectral image processing methods, it has become possible to quantitatively determine and classify various external characteristics of the color, shape, size, and surface texture of fruits and berries when assessing their quality, including for sorting [17]. The authors [18] conducted research on sorting apples by quality indicators (color, size, weight) in real-time using a vision system. The developed software processes four different images of a fruit in 0.5 s. The line is capable of sorting up to 15 apples per second, with sorting accuracy ranging from 73% to 96%.

A deep learning architecture based on convolutional neural networks (CNNs) was used to detect defective apples on a four-line fruit sorting machine running at 5 fruits per second. The CNN-based model demonstrated high performance on a test dataset with an accuracy, completeness, and specificity of 96.5%, 100%, and 92.9%, respectively [19]. The image processing method based on object counting on defective regions and the support vector machine (SVM) classifier showed less efficient performance, with accuracy, completeness, and specificity scores of 87.1%, 90.9%, and 83.3% respectively [20].

In [21], hyperspectral images of Pink Lady apples at three different harvest stages were analyzed to predict certain internal characteristics, such as firmness and soluble solid content (SSC). The results demonstrated that hyperspectral imaging coupled with artificial neural networks (ANNs) and decision tree (DT) methods proved to be more effective for predicting firmness, while DT and multiple linear regression (MLR) methods were more effective for SSC prediction.

In this study, spectral data ranging from 376 to 1011 nm were collected for all samples. Sample sets were partitioned using the physicochemical coeval distance method, and various spectral preprocessing methods were assessed by constructing a full-wavelength artificial neural network (ANN) model. Results indicated that the first-order difference + SG smoothing preprocessing method improved prediction accuracy. The CARS-ANN prediction model performed well with a root mean square error of prediction of 0.1150 and an R-value of 0.8675 for the prediction set [22].

A particular problem in the classification of fruit images is the similarity of the characteristics of stalks and calyxes with existing defects, which is a problem in sorting. The authors of the manuscript [23] proposed a method of apple defect detection based on laser-induced backscatter imaging and a convolutional neural network (CNN) algorithm. The CNN used provided a high recognition rate of 92.5% and higher accuracy than standard machine learning algorithms. In [24], we evaluated methods of hyperspectral image processing for apple defect detection. A two-band ratio was used to separate the stalk from the fruit. The authors took the average value of the visible and near-infrared regions to distinguish apple lesion areas. Hyperspectral analysis was performed with the hyperspectral range divided into three regions: Visible Spectrum (VIS), Near-Infrared (NIR), and Absorption Band (AB). The reflectance of each region was analyzed and Principal Component Analysis (PCA) was performed for each region.

The authors [25] applied a non-destructive approach to apple variety classification using a multi-channel hyperspectral imaging system incorporating illumination. Partial least squares discriminant analysis (PLS DA) models were developed for individual SR (spectral resolution) spectra and their combinations to evaluate their performance in variety classification. Results showed varying accuracy in classifying apple cultivars using single-SR spectra, with optimal SR spectra differing between spectral types. Using the method of binarisation and segmentation of hyperspectral images, it is possible to identify areas of apple fruit damage (to identify damaged skin) and images of objects having certain characteristics at certain wavelengths. These methods have been used to determine fruit quality, for example, to identify apples with damaged peel or signs of disease [26].

To efficiently process hyperspectral data, machine learning algorithms such as classification and clustering are used to automatically process large amounts of data and select objects with certain characteristics. For image binarisation, thresholding methods are used to select objects with a specific brightness or color. For image segmentation, clustering techniques are used to select groups of pixels with similar spectral characteristics. The majority of apple fruit diseases can be interpreted in the spectral regions from 380 to 1000 Nm [27].

A clustering method based on the spectral characteristics of pixels was used to identify fruits with disease signs. To obtain index images, mathematical formulas were obtained that allow us to calculate the spectral characteristics of objects based on the values of light intensity at different wavelengths [28,29].

To determine the chlorophyll content in plant tissue, wavelengths in the green and red spectral ranges are used, since chlorophyll has an absorption peak in these regions. Wavelengths in the infrared spectral range were used to detect skin lesions on the fruit surface because defects can change the intensity of reflected light in this region [30].

The authors [31] found that the use of index images can simplify the analysis of hyperspectral data and improve the accuracy of the results. The choice of wavelengths and the use of index images in hyperspectral image processing depends on the properties of objects and the equipment used [32].

An imaging spectrometer, which has a wider wavelength measurement range, was used to transform the image spectrum. The image in this case was categorized as multispectral or hyperspectral according to the spectral resolution. The authors of the study compared the performance based on the accuracy of the PCA, K-Means, and ISODATA; it was found that the overall accuracy of the K-Means algorithm was 78.34%, while the accuracy using the ISODATA algorithm was 81.77% [33].

The authors [34] used three effective tensors to compress hyperspectral images. The tensor decomposition method eliminates both spatial and spectral redundancy present in hyperspectral images. It is found that tensor decomposition using the HOSVD method outperforms PSNR and SSIM, with a high compression ratio compared to the other two methods.

In [35], we used a method of hyperspectral image processing, taking into account the spatial arrangement of pixels. The method jointly uses the results of pixel classification and the method of reference vectors and a set of contours obtained as a result of image clustering by the K-Means method. This method can also be used to solve the problem of image segmentation, which is valuable when processing hyperspectral data. As a result of using the combined method, the classification accuracy was increased to 96%, which is much more effective than using only the method of support vectors. Another machine learning method used in hyperspectral data processing and classification is SVM (support vector machine).

SVM can be an effective tool for the classification of complex data with non-linear relationships between features. In the manuscript [36], the problem of classification of hyperspectral remote sensing data was solved using the maximum likelihood (ML) method, spectral angle mapping (SAM), and support vector machine (SVM) method. The results obtained on the HYDICE sensor dataset showed that SVM is much more efficient than other traditional ML and SAM classifiers in terms of classification accuracy, computation time, and the stability of parameter settings.

Studies by the authors [37] have shown that the Minimum Distance classification method is quite simple and fast, but it may not give sufficiently accurate results if the classes have complex shapes or overlap with each other. In such cases it may be more effective to use more complex classification methods, such as the support vector method. The authors of [38] apply the Binary Encoding classification method for hyper-spectral image classification. An original Binary Encoding classification algorithm is presented, in which regions rather than single pixels were taken as the basis, and information about the size, shape, and height of each individual segment was included in the classification. The results of using this method showed that it provided insignificant accuracy gains over SVM classification. Another method of the supervised classification of hyperspectral images is Maximum Likelihood Classification. Spectral subgroups are formed from the full spectral dataset by analyzing the global correlation between spectral bands. Scientific studies show that optimal sizes of such subgroups can provide a balance between classification accuracy, processing time, and the availability of training pixels [39,40,41].

The detection of the surface damage of fruits at early stages, before storage, plays an important role in their automated sorting and the improvement of product quality after removal from storage. The choice of the best method of fruit optical visualization and efficient image processing is an actual subject of scientific discussions and research [42,43].

The results of the study confirm the relevance of using hyperspectral analysis for fruit sorting, as also highlighted in recent works such as [44], which demonstrates a maturity assessment of Achacha fruit, [45], which explores sensitivity analysis in digital twin models for air pollution, and [46], which uses hyperspectral analysis for sustainable waste recycling.

The aim of the study is to develop a device capable of evaluating the damage of apple fruits on a conveyor using hyperspectral image analysis, as well as evaluating the effectiveness of methods for the classification and recognition of hyperspectral images of diseased apple fruits (Figure 1).

The main contributions of the conducted research are as follows:

• An analysis of hyperspectral data classification methods for recognizing apple damage was carried out, demonstrating performance comparable to methods used by experts.
• Hyperparameter optimization was performed for machine learning methods such as SVM, K-Means, and others, which improved the accuracy of damage recognition caused by various apple diseases.
• The potential of using hyperspectral cubes for more accurate recognition and classification of apple diseases was investigated, confirming the prospects of this approach for application in the agro-industrial complex.
• Differences in the spectral characteristics of apples affected by diseases such as fire blight, scab, and rot were analyzed, allowing for more accurate damage recognition.
• The advantages of using hyperspectral data in the agro-industrial complex were identified, demonstrating that these data have high potential for application in automatic fruit sorting systems, opening up new opportunities for their implementation in agriculture.
• To compare the efficiency of different classification methods, both supervised methods (SVM, Minimum Distance, Maximum Likelihood) and unsupervised methods (IsoData, K-Means) were used, which allowed us to identify the most efficient approaches to classifying hyperspectral data.
• A specially designed hyperspectral image collection rig was used during the study, ensuring high data quality and comparability with real conditions on sorting lines.

2. Materials and Methods

2.1. Characteristics of the Research Subject, Apple Fruit

Fruit of the apple varieties Honeycrisp (Macoun × Honeygold)—n = 200, Gala (Kidd’s Or-ange × Golden Delicious)—n = 200, and Jonagold (Jonathan × Golden Delicious)—n = 200 were used as research objects. The sample included fruits affected by the following diseases: sunburn, scab, and rot. Apples of the Honey Crisp variety were 7.5–9 cm (2.95–3.54 in) in diameter and about 8.4–10.2 cm (3.31–4.02 in) in height. Apples of the Gala variety had a diameter of 6.2–7.5 cm (2.44–2.95 in) and a height of 6.6–7.4 cm (2.60–2.91 in). Apples of the Jonagold variety were 7.5–8.9 cm (2.95–3.50 in) in diameter and 7.5–8.2 cm (2.95–3.23 in) in height. The extent of apple lesions ranged from 10% to 50% of the total apple area. The apples were stored in refrigerated chambers with a controlled atmosphere, the humidity level was 90%, and the temperature re-mode ranged from −2 °C to +7 °C until the time of research.

2.2. Real-Time PCR

Primers specific for the pathogens of yavlok fruits were used: Venturia inaequalis, Botrytis cinerea, Penicillium expansum, and Erwinia amylovora (Table S1) [47,48,49,50]. The primers were synthesized by holding “Eurogen” (Moscow, Russia). The reaction mixture was prepared by mixing 5 μL of qPCRmix-HS SYBR (“Eurogen”, Moscow, Russia) with a pair of target primers (1 μL of each), 1 μL of DNA template solution (1.3 × 10² ng/mL), and 23 μL of Milli-Q water. The real-time PCR reaction was carried out in a Thermo Applied Biosystems QuantStudio 5 Real-Time amplifier (Waltham, MA, USA); the conditions have been described in detail previously [51]. Fluorescence intensity measurements were performed at the end of the 72 °C cycle. The method for obtaining Ct values, standard curves, and corresponding correlation coefficients (R²) was described previously [52]. As a negative control, 2 µL of Milli-Q water was added to the reaction mixture instead of the DNA.

2.3. Measurement of Fluorescence of Control and Infected Apples

The fluorescence of control and damaged apple surface areas was studied using a Jasco FP-8300 spectrofluorometer (Tokyo, Japan). Spectral portraits of apple parts were measured in a special cuvette for dry samples. All parameters were the same for all measurements and were chosen so that the intensity peaks corresponded to approximately one fifth of the entire intensity range of the device [53]. In addition, the width of the “excitation band” was maximized to minimize the influence of the geometry of the apple parts in the cuvette. Each sample was measured at least five times. Measurements were carried out at room temperature of approximately 22 °C. The experimental techniques used in the measurements have been described in detail elsewhere [54].

2.4. Stand for Hyperspectral Measurements

To collect hyperspectral data, a BaySpec OCI-F hyperspectrometer (San Jose, CA, USA) with a near-infrared (NIR) lens with a focal length of f = 16 mm and coverage of 24° FOV (San Jose, CA, USA) was used and mounted on the stand suspension. The stand suspension has the ability to move horizontally and vertically at a predetermined speed. A stepper bipolar motors model PL57H110-B8 (Kanata, ON, Canada) with ball screw transmission DSG1605/1610 (Kanata, ON, Canada) and a bearing support model BF12 (Kanata, ON, Canada) were used as drives. A table with rubber rollers was used as a mechanical system to transport and rotate the apples during hyperspectral image acquisition. An Arduino Micro controller board based on the ATmega32u4 (Manila, Philippines) with DM556 drivers (Leadshine Technology Co., Ltd., Hangzhou, China) and LRS-350-36 power supply (Leadshine Technology Co., Ltd., Hangzhou, China) was used to control the suspension movement. Four tungsten halogen lamps of the Camelion GU10 model (Camelion, Hangzhou, China) with a power of 35 W were used for illumination of apple tree fruits during their movement on the table, which are fixed on the stand suspension. For data analysis, we used a computing platform based an Intel Core i9-10900X (Santa Clara, CA, USA) processor with 10 cores, 20 virtual threads, two NVIDIA Ge-Force RTX 2080 Ti graphics cards (Santa Clara, CA, USA), a GIGABYTE X299 UD4 Pro motherboard (Taipei, Taiwan), Intel PCI-E 1Tb 660P SSD (Santa Clara, CA, USA), and a Kingston DDR4 DIMM 32 GB memory module (Fountain Valley, CA, USA).

2.5. Hyperspectral Data Treatment

SpecGrabber Pro software (V.1.1) was used to obtain the original hyperspectral images. The experiments used 8-bit data (8 bits/channel), including high spatial resolution RGB image data acquired using Sensor 1 and hyperspectral data acquired using Sensor 2 of the BaySpec OCI-F hyperspectrometer. Shooting parameters such as exposure time (exposure) and gain were determined through a series of preliminary experiments, during which the optimal values of these parameters were empirically selected. Based on the results of these experiments, it was decided to use auto exposure and set the gain to very low (VL) to achieve the best hyperspectral image quality. To calibrate the hyperspectrometer, a standard sample with 100% reflection, representing a white body of A4 size, was scanned sequentially, and the values of currents in a dark background were recorded to perform calibration. The speed of saving images was 50 FPS and the speed of suspension movement was 0.05 m/s. As a result of these studies, a set of hyperspectral images of the research objects was obtained. For each pixel in the acquired hyperspectral image, the spectral response was recorded in the spectral range from 400 to 1000 nm with a resolution within 5–7 nm and a resolution of 512 pixels in the horizontal direction per image line. BaySpec CubeCreator 2100 software (San Jose, CA, USA) was used to process the raw images acquired with the BaySpec OCI-F hyperspectrometer and convert them into hyperspectral cubes. The software allowed us to group pixels with similar spectra based on Pearson’s linear correlation algorithm when the correlation value exceeded the classification threshold defined in the parameters. Gelion Beta 1.0 software (Gelion Team, Moscow, Russia) was used to control the regions of interest and analyse the spectral characteristics of given classes during annotation.

The analysis showed that the spectral characteristics of sunburn lesions show an increase in reflectance in the near-infrared (NIR) range because the damaged areas reflect more infrared radiation due to the presence of damaged tissue. The spectral characteristics of a scab include increased reflectance in the visible range, especially in the range corresponding to the color of the spots. The rot spectral characteristics include increased reflectance in the near-infrared range due to changes in the chemical composition and structure of the affected areas.

As a result of the conducted studies, spectral characteristics of the given classes of apple fruit lesions of Honey Crisp, Gala, and Jonagold varieties were obtained.

The annotation of the obtained hyperspectral images and identification of regions of interest (ROIs), were performed by experts using the ENVI 5.7 software package (ITT Visual Information Solutions, CO, USA) using tools for drawing contours around objects or selecting pixels belonging to each class. The expert group included specialists in both bioengineering and crop breeding. The classes chosen were “Apple_red” for intact red apple regions, “Apple_green” for intact green apple regions, “Apple_gray” for intact gray apple regions, “Branch” for apple stalk, “Sunburn” for apple regions affected by sunburn, “Scrab” for apple regions affected by scab, “Rot” for apple regions affected by rot, and “Background” for background regions.

Labeled datasets were created and used for training and testing classification and recognition methods on hyperspectral images by various machine learning algorithms. The dataset of apple fruits of three varieties was divided into training samples (70% of data) and test samples (30% of data). The training sample was used to train the models and the test sample was used to evaluate the performance of the models on new previously unused data. The fruit images were divided into these two samples (420 and 180 pieces) based on their description and characteristics. The uniform distribution of images across apple varieties was taken into account to ensure the representativeness of each variety in the training and test samples.

Each image pixel in the dataset is represented by a feature vector that including spectral reflectance coefficients. The classifiers were trained using all image pixels as training data. These results were compared with expert annotations, which served as a benchmark for evaluating the accuracy of the classifiers.

2.6. Classification and Recognition Methods

As a result of analyzing the research on the application of classification and recognition methods for hyperspectral images, the most effective methods have been selected to evaluate their performance, as discussed in the Introduction section. For this purpose, supervised classification methods such as Binary Encoding, Maximum Likelihood, Minimum Distance, Parallelepiped, support vector machine, and unsupervised classification methods such as IsoData and K-Means Classification were used. The difference between supervised and unsupervised methods is that the former requires examples of labeled images, while the latter performs classification based only on the existing image (without labeling).

The Binary Encoding classification method allows for the division of image pixels represented by spectral vectors into classes using pre-selected threshold values for pixel separation. The Distance classification method calculates distances between the spectral vector of a pixel and the centroids of classes, assigning to a pixel the class whose centroid is at the minimum distance. In this method, each pixel is classified based on its spectral characteristics and similarity to predefined classes. The Maximum Likelihood classification method based on a statistical approach and maximum likelihood allows us to determine the class to which each pixel in the hyperspectral image belongs with the highest probability. The Minimum Distance method in hyperspectral image classification allows for the assignment of a pixel to a particular class based on the minimum distance between its spectral characteristics and the average spectral characteristics of each class. The Parallelepiped classification method allows for the classification of each pixel in the image based on the comparison of its spectral characteristics with predefined intervals (Parallelepiped) for each class. A pixel is considered to belong to a class if its spectral characteristics fall into the interval of this class. The support vector machine classification method is a machine learning method that separates pixels in hyperspectral images into different classes, using a hyperplane and a kernel function to determine which side of the hyperplane an image pixel falls on. The IsoData Classification method allows iterative clustering of pixels into groups (clusters) based on their spectral characteristics using spectral proximity criteria and automatically separates hyperspectral data into several classes. The K-Means classification method allows for clustering, which divides pixels in an image into groups (clusters) of similar spectral characteristics by minimizing the variance within a cluster and maximizing the differences between clusters, allowing objects to be segmented and classified based on their spectral information. These classification and recognition methods show significant potential for accurate and efficient processing of hyperspectral data, which confirms their importance in the field of scientific research. It should be noted that unlike the widely used RGB optical images, where each pixel is described by a vector of three coordinates, the number of channels in BaySpec OCI-F hyperspectral images reaches 240, which allows one to obtain better results during training.

2.7. Performance Evaluation Method

It should be noted that unlike the widely used RGB optical images, where each pixel is described by a vector of three coordinates, the number of channels in BaySpec OCI-F hyperspectral images reaches 240, which allows one to obtain better results during training.

To evaluate the classification accuracy of each class, reference masks are created on the image. Since the marked image has no more than N possible brightness values, TIFF images obtained in ENVI software (V.5.0) can be used for quality assessment. To obtain the brightness of pixels of each class in the reference image, a transformation of the images into a two-dimensional representation was performed using Formula (1):

$$I_{gs}=0.299I_r+0.587I_g+0.114I_b,$$

(1)

where I_gs—image normalized by three channels; I_r, I_g, I_b—brightness in TIFF image in red, green, and blue color channels respectively.

This allows you to have a reference brightness for each class and apply the ARI coefficient only to the areas of interest. Binary masks were then created for each class: a brightness value of 255 was set for the background and the object was left with the brightness value it had in the reference image. This method allowed the visual separation and identification of classes in the image and was used to assess the accuracy of classification and the identification of specific objects or regions of interest.

However, the visual analysis of the mask data revealed boundary artifacts and errors in the background region of the image. To eliminate this problem, masks were created not on the entire reference image, but only on regions containing pixels of the same class. Since the K-Means and ISODATA algorithms return clusters instead of classes, the ARI method was used to evaluate their accuracy. The most popular cluster in the domain of interest was identified and compared to a benchmark. To evaluate the accuracy in the case where all pixels of each class N_k are known, we compare how many pixels from this region the model assigned to this class M_k and calculate the accuracy using Formula (2):

$$A_{c{c}_k}=M_k/N_k$$

(2)

where k—particular class.

After this evaluation, the metrics and the number of pixels in each given class are obtained. Using the ARI Formula (3) for each class, the accuracy is determined:

$$ARI=2·(\mathrm{a}\mathrm{d}-\mathrm{e}\mathrm{d})/(\mathrm{a}\mathrm{d}+\mathrm{e}\mathrm{d})$$

(3)

where ad (Agreement on Data)—number of pairs of objects that are classified identically both in the model (as clusters) and in the benchmark (as classes); ed (Expected Agreement on Data)—expected number of pairs of objects that would be classified identically at random.

Thus, ARI was calculated for each region of interest using clustering performed by the K-Means and ISODATA algorithms, and comparison of the results with reference classes to obtain information on the similarity between clusters and classes was performed. Figure 2 shows the architecture diagram of the proposed approach.

The general structure of the study includes collecting and preprocessing hyperspectral images, calibrating the equipment, splitting the data into training and test sets, applying preprocessing methods, implementing machine learning for disease classification, evaluating the effectiveness of the method using ARI, and checking the results by comparing with visual assessments and expert data. The developed device for obtaining hyperspectral images and automated apple sorting (Figure 3) structurally consists of a sensor (pos. 1, Figure 3); an optical module (pos. 2, Figure 3); a module for lighting the test sample; a conveyor belt (pos. 4, Figure 3); a short filter, an objective lens, and an adjustable slit (pos. 5–7, Figure 3); a CCD matrix with a controller (pos. 8–9, Figure 3); a light-conducting flux, a light prism, and a reflector (pos. 10–13, Figure 3); a package of fluorescent lamps (pos. 14, Figure 3); a long-pass light filter (pos. 15, Figure 3); a micrometer drive (pos. 16, Figure 3); and a conveyor belt (pos. 17, Figure 3). Image preprocessing included noise reduction and spectrum normalization techniques, such as smoothing with a first-derivative filter. These steps ensure the stability of the model and improve classification accuracy, especially when processing hyperspectral data. The imaging parameters were selected to achieve optimal hyperspectral image quality. The exposure time was set to 200 ms, the gain was 1.5, and the lighting angle was 45°, which ensures uniform illumination and high contrast when shooting the surface of apples.

3. Results and Discussion

The objects of the study were the fruits of the HoneyCrisp (Macoun = Honey gold), Gala (Kidd’s Orange × Golden Delicious), and Jonagold (Jonathan × Golden Delicious) apple varieties. The fruits were affected by scab, rot, and sunburn. Representative photographs of the fruits used in the experiments are shown in Figure 4. Sunburn on apples appeared as dark spots or layers of blackened, dry skin. Apples affected by apple scab have dark spots on the skin of the fruit and the texture of the spots is uneven or rough. Damage from apple rot is manifested in the form of rotting and the deterioration of the pulp of the fruit, dark spots, and superficial or internal rotten areas.

To confirm sunburn, scab, and rot, in addition to the expert method, 3D fluorometry and real-time PCR were used. The appearance of sunburn on the surface of apples is usually caused by prolonged exposure to sunlight [55]. For this reason, sunburn cannot be identified using real-time PCR, but it can be characterized using fluorescence spectroscopy. Apple scab is usually caused by a fungal disease known as Venturia inaequalis and is one of the common diseases of apples [56]. Apple rot is usually caused by the fungi Botrytis cinerea and Penicillium expansum, as well as the bacteria Erwinia amylovora [57]. The causative agents of scab and rot on apple fruits were identified using the real-time PCR method (

Table 1. Real-time PCR results for identification of scab and rot pathogens on apples.

Pathogens	RT-PCR, Seed Infection Rate, Ct
	Scab		Rot
	+	–	+	–
Venturia inaequalis	17	>40	>40	>40
Botrytis cinerea	>40	>40	15	>40
Penicillium expansum	>40	>40	22	>40
Erwinia amylovora	>40	>40	>40	>40

). It was shown that the apple sample identified by the expert as a scab contained the presence of the pathogen Venturia inaequalis, which is characteristic of scab disease. At the same time, the sample of the same apple taken from another, visually healthy location did not contain the pathogen Venturia inaequalis (Ct > 40). It was shown that the apple sample identified by the expert as rot contained the presence of the pathogens Botrytis cinerea and Penicillium expansum. The presence of these pathogens is characteristic of apple rot. The presence of another pathogen characteristic of apple rot, Erwinia amylovora, was not detected. At the same time, the samples taken from another, visually healthy location did not contain the presence of pathogens (Ct > 40). To minimize the influence of external factors on the results of hyperspectral analysis, apples were stored in controlled conditions at 90% humidity and temperatures from −2 °C to +7 °C before the start of shooting. This made it possible to standardize the conditions and improve the reproducibility of the results.

In parallel, the samples were identified using optical methods. The fluorescence spectra of healthy and damaged areas on the apple surfaces were studied (Figure 5). The following fluorescence intensity maxima were found in healthy apples: 2782 a.u. at λ_ex = 264 nm, λ_em = 329 nm; 1024 a.u. at λ_ex = 348 nm, λ_em = 403 nm; 540 a.u. at λ_ex = 434 nm, λ_em = 683 nm; 3245 540 a.u. at λ_ex = 670 nm, λ_em = 685 nm. The following fluorescence intensity maxima were found in apples affected by rot: 711 a.u. at λ_ex = 266 nm, λ_em = 328 nm; 545 a.u. at λ_ex = 348 nm, λ_em = 404 nm; 558 a.u. at λ_ex = 668 nm, λ_em = 683 nm. The following fluorescence intensity maxima were found in apples affected by scab: 1527 a.u. at λ_ex = 274 nm, λ_em = 297 nm; 644 a.u. at λ_ex = 348 nm, λ_em = 423 nm; 405 a.u. at λ_ex = 474 nm, λ_em = 684 nm; 420 a.u. at λ_ex = 438 nm, λ_em = 685 nm; 1813 a.u. at λ_ex = 672 nm, λ_em = 687 nm. The following fluorescence intensity maxima were found in apples affected by sunburn: 1006 a.u. at λ_ex = 274 nm, λ_em = 297 nm; 154 a.u. at λ_ex = 346 nm, λ_em = 424 nm. The characteristic dominant fluorescence at λ_ex = 264 nm, λ_em = 329 nm, mediated by aromatic amino acids [58], is observed only in healthy and rotten apples. In the case of apples affected by scan and sunburn, a shift in the dominant fluorescence intensity to a short wavelength region is observed. This is usually associated with oxidative damage to aromatic amino acids.

On 3D maps of the fluorescence intensity of all apples, there is a maximum with coordinates at λ_ex = 340 nm, λ_em = 445 nm. The presence of such a maximum is usually associated with a fairly large area, which is formed as a result of the combined fluorescence of a large number of substances (phenolic acids (hydroxycinnamic), ferulic acids, coumaric acids, pteridine compounds (folic acid, neopterin, etc.), chitin, coumarin, chlorogenic acid, NADPH, NADH, cellulose, lignin, suberin, lipofuscin, flavonoids, alkaloids, sporopollein, terpenoids, and mycotoxins (deoxynivalenol, nivalenol, zearalenone, and alternariol)) [59]. Usually, a significant increase in the size of this area is associated with fungal or bacterial contamination [60]. It should be noted that this area is the largest in apples affected by rot and scab. A separate group of areas with an emission of about 680 nm is apparently associated with photosynthetic pigments, or more precisely with chlorophyll [61]. It should be noted that the diversity of areas of fluorescence of photosynthetic pigments is observed only in the case of scab and healthy apples. In the case of rot and sunburn, chlorophyll fluorescence is 5–6 times lower than in healthy apples. Thus, using several independent methods (expert assessment, PCR, fluorometry), it was shown that healthy apples and apples affected by rot, scab, and sunburn participated in the experiments. All apples were sorted into groups and labeled. After characterizing the apples, we proceeded to the implementation of the main goal set in this work. Figure 6 shows a representation of the data obtained using a hyperspectral camera. The frame has manually selected areas and their characteristic spectra: 1—background, area to the left of the apple; 2—petiole area; 3—area along the apple contour, the upper left edge of the apple; 4—area of the apple with red skin; 5—area of the apple with green skin; 6—area affected by rot, dark spot in the center. Thus, with the help of hyperspectral photography, it is possible to identify various typical areas on apples. The process of selecting and verifying Regions of Interest (ROIs) in this study was performed manually and included the following steps. In each image, areas with damage or defects were manually highlighted based on visual characteristics such as color changes, texture, and the surface structure of the fruits, allowing the precise identification of zones for further spectral analysis. Each ROI was carefully examined to confirm its correspondence to the damaged area, with verification by experts to ensure accuracy and consistency. This process was repeated for each image, particularly in cases with subtle defects or minimal texture deviations. The manual selection and verification of ROIs ensured high accuracy and alignment with the study’s requirements, enabling precise analysis of damages and their spectral characteristics. To establish the basic spectral characteristics, a control group of healthy apples without signs of damage was used. This made it possible to better distinguish between affected and healthy fruits and increase the accuracy of damage identification.

Our team created a sorting line unit that allows us to identify apples that have damage from healthy fruits (Figure 7). The sorting line unit consisted of stepper bipolar motors, a ball screw transmission, a transmission, a table with rubber rollers, light sources, a hyperspectral camera, and a control unit. It was shown above that in static mode, the approach we proposed effectively allows us to distinguish between various types of fruit damage. When moving from static conditions to a sorting mechanism, questions naturally arise about the maximum speeds at which it is possible to use the hyperspectral method for detecting apple damage [62].

To answer the question about the maximum speeds at which the hyperspectral method for detecting apple damage can be used, the effect of the rotation speed of the rubber rollers of the table on the quality of hyperspectral imaging was studied. The quality of hyperspectral imaging was understood as the ratio of frames suitable for analysis to the total number of frames. Figure 8a shows examples of both suitable and unsuitable frames taken of the same apple at different rotation speeds of the rubber rollers. Figure 8b shows the dependence of the efficiency of hyperspectral imaging on the rotation frequency of the rubber rollers of the sorting line unit. It was found that the maximum efficiency of hyperspectral imaging is observed at rotation frequencies of up to 6 Hz. With an increase in the rotation frequency of the rubber rollers, the efficiency of hyperspectral imaging significantly decreases. In all subsequent experiments, a rotation frequency of the rubber rollers of 6 Hz was used.

To improve the stability of the quality of automated fruit sorting, the optimal parameters of the sorting section with an identification module above the roller table were experimentally established: a carriage movement speed of 0.01–0.09 m/s, a light level (four halogen lamps) of 2600 lk, and an apple scan time of 1–30 s.

Figure 9 shows the dependence of the roller rotation frequency (b), which is regulated using a three-phase current frequency converter (a). As seen in the figure, the optimal values of the roller rotation frequency (c), at which the damage rate of apples is less than 5%, are up to 6 Hz. At this frequency, the carriage with the camera fixed on it also moves parallel to the working surface, at a speed of up to 0.09 m/s, depending on the required scanning time (from 1 to 30 s) and the flow of apples. At a frequency greater than 6 Hz, the apples begin to move uncontrollably along the working surface, and, reaching a frequency of 25 Hz, they fly off the conveyor.

The fruit of the apple varieties Honey Crisp (Macoun × Honeygold) n = 200, Gala (Kidds Or-ange × Golden Delicious) n = 200, and Jonagold (Jonathan × Golden Delicious) n = 200 were used as research objects. The data were divided into training and test sets. 70% of the data (420 images) were used for training the models, while 30% (180 images) were used for testing. The training set included an even distribution of images by variety to ensure the representativeness of the data for each model. We used custom datasets containing images collected in laboratory conditions using a dedicated setup. This allowed for a more comprehensive investigation and validation of the proposed methods on data close to real-world conditions, ensuring their practical applicability. The sample included fruits affected by the following diseases: sunburn, scab, rot. An example of visualization of apple fruits, as a result of recognition and classification by different methods is presented in Figure 10.

To ensure the accuracy and reliability of the data, all experiments were carried out using 10-fold cross-validation. Each method (supervised—Binary Encoding, Maximum Likelihood, Minimum Distance, Parallelepiped, SVM, as well as unsupervised—IsoDATA and K-Means) was tested 10 times on each dataset to reduce random variations and increase statistical reliability. To improve the perception of the results, the medians and standard deviation for each method were added to the final tables in the text of the article in order to confirm the statistical reliability and stability of the obtained metrics. Hyperparameter settings were performed for each method in order to optimize their performance on hyperspectral data. For the Binary Encoding method, the threshold value for pixel separation was established based on an analysis of the intensity distribution, while the threshold was adapted for each class in the range of 0.5 to 0.7 to achieve maximum classification accuracy. For the Maximum Likelihood method, a covariance matrix was used for each class, and a priori probabilities were adjusted depending on the proportion of classes in the sample, which ensured a balance between large and small classes. For the Minimum Distance method, Euclidean distance was used as a metric, with the setting of the distance parameter for each class group to minimize classification errors in places where classes overlapped. For the Parallelepiped method, the boundaries for each class were determined based on the spectral characteristics of pixels. The hyperparameters were adjusted by varying the width of the parallelepipeds to minimize false classifications, especially at class boundaries. For the SVM method, the hyperparameters included the core selection (RBF core), the regularization parameter (C) configured in the range of 0.1 to 10 in 0.1 increments, and the gamma coefficient (γ), which was configured for optimal class separation. For the IsoData (unsupervised) method: The main hyperparameters of the IsoData (unsupervised) method included a minimum number of points for each cluster, configured within 2–5 points, and a cluster fusion threshold set to 0.95 to ensure the accurate separation of spectral data. The number of clusters for the K-Means method ranged from 3 to 6, depending on the complexity of the spectral characteristics of the data. The algorithm used the K-Means++ initialization method to improve the clustering quality. Each of these hyperparameters has been optimized using cross-validation on a training sample to ensure that the selected values minimize error and provide stable results. Based on the analysis results of different methods on recognition and the classification of the test samples of apple fruits affected by sunburn, it was determined that the SVM method shows the highest mean ARI (0.811), indicating its relatively high performance on average in all classes with sunburn-affected apple fruits (

Table 2. Results of quality analysis of methods for classification and recognition of apple tree fruits affected by sunburn test sample.

Method	Background	Branch	Sunburn	Apple Red	Apple Green	Apple Gray	Average ARI	Median Accuracy (%)	Standard Deviation (%)
Supervised Classification
Binary Encoding	0.589	0.650	0.681	0.690	0.680	0.650	0.657	65.7	±3.2
Maximum Likelihood	0.604	0.750	0.704	0.790	0.780	0.740	0.728	73.7	±2.8
Minimum Distance	0.327	0.753	0.732	0.712	0.765	0.780	0.678	67.8	±4.5
Parallelepiped	0.327	0.650	0.680	0.690	0.680	0.650	0.613	61.3	±3.1
SVM	0.665	0.835	0.789	0.878	0.869	0.829	0.811	81.1	±2.5
Unsupervised Classification
IsoData	0.535	0.613	0.623	0.644	0.637	0.608	0.610	60.7	±3.3
K-Means	0.714	0.824	0.667	0.478	0.812	0.546	0.674	67.4	±4.0

).

From the results of the analysis of different methods for the recognition and classification of apple fruits affected by scab, it was determined that the SVM and K-Means methods showed the highest mean ARI (0.760) and ARI (0.657), respectively (

Table 3. Results of quality analysis of methods for classification and recognition of apple tree fruits affected by scab test sample.

Method	Background	Branch	Scab	Apple Red	Apple Green	Average ARI	Median Accuracy (%)	Standard Deviation (%)
Supervised Classification
Binary Encoding	0.474	0.532	0.506	0.540	0.584	0.527	52.7	±3.8
Maximum Likelihood	0.356	0.612	0.581	0.621	0.671	0.568	56.8	±3.1
Minimum Distance	0.438	0.641	0.651	0.696	0.752	0.636	63.6	±4.2
Parallelepiped	0.449	0.490	0.466	0.526	0.538	0.494	49.4	±3.0
SVM	0.523	0.759	0.818	0.755	0.945	0.760	76.0	±2.6
Unsupervised Classification
IsoData	0.598	0.578	0.457	0.618	0.528	0.556	55.6	±3.9
K-Means	0.640	0.643	0.786	0.605	0.610	0.657	65.7	±3.5

).

Based on the determined ARI values, the K-Means method and SVM methods were the best options to recognize and classify apple fruit affected by rot (rot class) in the used dataset (

Table 4. Results of quality analysis of methods for classification and recognition of apple tree fruits affected by rot test sampling.

Method	Background	Branch	Rot	Apple Red	Apple Gray	Average ARI	Median Accuracy (%)	Standard Deviation (%)
Supervised Classification
Binary Encoding	0.602	0.675	0.674	0.766	0.682	0.680	68.0	±2.9
Maximum Likelihood	0.726	0.736	0.715	0.800	0.780	0.751	75.1	±3.0
Minimum Distance	0.307	0.812	0.831	0.776	0.865	0.718	71.8	±4.5
Parallelepiped	0.512	0.623	0.736	0.806	0.723	0.680	68.0	±3.8
SVM	0.862	0.854	0.854	0.870	0.863	0.861	86.1	±1.9
Unsupervised Classification
IsoData	0.738	0.685	0.574	0.617	0.584	0.640	63.9	±3.7
K-Means	0.580	0.687	0.840	0.722	0.806	0.727	72.7	±3.2

). The SVM method in the used test sample has a higher average ARI value (0.861).

The most sensitive and accurate method was SVM. The average ARI was 25.7% higher on average for all classes. At the same time, the K-Means algorithm is characterized by high metrics for some classes but low ones for others, which makes it less preferable in general, but a good option for scab and rot disease detection.

In this study, different approaches to hyperspectral reflectance data processing were tested for the detection of damaged apple fruit sections using different methods. As a result of these studies, it was found that non-invasive recognition of apple fruit damage using hyperspectral images is possible with an accuracy comparable to that of human experts. In this research work, a significant variation in the ARI values obtained when classifying the regions associated with the “Apple_red” and “Apple_green” classes is observed. This observation is explained by the spectral characteristics of these areas, which in most cases show insignificant differences between them.

The results of this study confirm the high potential of hyperspectral data for detecting and classifying apple diseases and lesions. Machine learning methods such as SVM and K-Means have proven to be effective in this task. Hyperspectral data have significant potential to accurately classify and recognize diseases and lesions on apples. These data have a wealth of information on the spectral characteristics of the fruit that can be used to detect even subtle changes in fruit condition. Further research could focus on improving classification accuracy and extending the application of hyperspectral data to other types of apple fruit diseases. Continuing the research on hyperspectral analysis, it is necessary to emphasize the importance of using hyperspectral cubes to achieve high efficiency in the detection and classification of diseases and lesions on apple fruit. Multidimensional data structures, where each element (pixel) contains spectral information covering different wavelengths in the visible and infrared spectral bands, allow a deeper analysis of fruit characteristics. It is important to emphasize that the processing and analysis of the vast amounts of data collected using hyperspectrometers requires significant computational resources and specialized software. This is due not only to the large volume of data, but also to the high dimensionality of hyperspectral cubes, where each element is a multidimensional vector of spectral values. The effective implementation of methods for analyzing and classifying hyperspectral data becomes a critical task for further research in this area. Such methods should take into account complex data structure, specific spectral characteristics, and peculiarities of different classes of objects. The need to develop high-performance algorithms and software solutions for the processing and interpretation of hyperspectral data is a key component of research work in this area. The proposed method can be adapted for other types of fruits, such as pears or kiwis, taking into account their unique spectral characteristics. Training the model on data from other fruits can provide a reliable detection of lesions on various fruits, assuming calibration and an optimal spectrum selection.

The study had a number of limitations, such as the sensitivity of hyperspectral imaging to changes in lighting conditions and the speed of the sorting line. For example, images taken at a high line speed were sometimes blurred, and changes in lighting intensity affected image quality and classification accuracy. Preliminary analysis has shown that changes in illumination intensity and spectral resolution significantly affect image quality and classification performance; further study will help optimize these parameters for better results.

We plan to conduct a sensitivity analysis on key parameters of machine learning algorithms, such as the kernel parameters in SVM and the number of clusters in K-Means. This will help optimize parameter selection and improve defect detection accuracy in agricultural applications. In future research, it is necessary to eliminate these limitations by developing more stable image processing methods that can adapt to dynamic conditions on the sorting line. In future research, we plan to conduct a comparative analysis to evaluate how differences in the extent of disease lesions affect classification accuracy. Relationships between lesion severity at different disease stages will be established. This approach will allow for the assessment and improvement of the model’s robustness and accuracy when dealing with varying degrees of lesions, ultimately enhancing its accuracy in agricultural applications.

Further research is also recommended to study the adaptation of the model to other types of fruits and damages, which will expand the scope of hyperspectral analysis in agriculture. Technical recommendations for specialists include the calibration of lighting systems to stabilizing image quality and optimizing line speed to avoid the blurring of images.

4. Conclusions

Thus, in order to improve the sustainability of the quality of automated fruit sorting by analyzing hyperspectral images and applying machine learning methods, a device has been developed that can assess the damage of apple fruits on a conveyor using hyperspectral image analysis. The efficiency of classification methods and capturing hyperspectral images of damaged apple fruits was evaluated. The general innovativeness of the work lies in the fact that the device and software we have developed allow farmers to reduce the manual labor of sorting apples. In other words, our work was devoted to the automation of labor mechanization in agriculture and increasing the sustainability of fruit production and crop quality prior to storage.

As a result, we obtained results on the recognition and classification of hyperspectral images, as well as a study of the possibility of analyzing hyperspectral and RGB images with further processing of spectral information, using supervised and unsupervised classification methods.

Analysis shows that the SVM method performed well in the task of classifying hyperspectral images of apple fruits affected by various diseases. For three classes, such as sunburn (0.789), scab (0.818), and rot (0.854), the SVM method showed the highest average ARI (Adjusted Rand Index). This indicates its ability to effectively separate and classify different classes of phenomena. The K-Means method also showed good performance for detecting the scab (0.786) and rot (0.84) classes, but did not show high performance for some other classes. This makes it less advantageous for the general classification of hyperspectral images.

The results of this study confirm the high potential of hyperspectral data for apple disease classification. In modern machine learning approaches, it is becoming evident that the integration of spectral data about the objects in question can significantly improve the efficiency of classification and analysis algorithms.

Modern machine learning algorithms focused on RGB image analysis show the potential to improve their performance when combined with spectral data describing the objects in question (apples). For optimal implementation, this measure avoids capturing images in the areas of interest using multiple strategies of selected spectral bands. The integration of supervised classification methods using pre-localized data and unsupervised classification to refine clusters and groups in the data provides the opportunity to further improve the detection and classification accuracy. This approach facilitates identification by providing new classes or patterns that could be sufficiently taken into account in the predefined classification schemes. Future research can be based on a higher classification accuracy, using deep learning models like YOLOv8-seg with hyperspectral data. In addition, expanding the research to other fruit species and diseases, as well as improving the data processing and analysis processes, is also of interest to further improve the efficiency of sorting lines in the agro-industrial complex.